DeLTA seminar by Rasmus Pagh
Speaker
Rasmus Pagh, professor in the Algorithms and Complexity section at Department Computer Science, University of Copenhagen.
Title
Efficient differential privacy in the shuffle model
Abstract
Differential privacy is a formal constraint on randomized mechanisms for privately releasing results of computations. In recent years, differentially private algorithms for machine learning have been developed and made available, for example, in TensorFlow/privacy and the Opacus library for PyTorch. Using such algorithms ensures that the presence or absence of a single data record in a database does not significantly affect the distribution of the model produced.
Concurrently, the area of federated learning has explored how to carry out machine learning in settings where training data is distributed among users and never shared. The shuffle model is a scalable framework for collecting data in federated settings, hiding the origin of each data item. This talk surveys recent advances in using the shuffle model to implement differentially private mechanisms. In many cases, the privacy-utility trade-off is close to that of a curator model, in which all data is transmitted to a central curator who performs the computation.
Based on joint work with Ghazi, Golowich, Manurangsi, Kumar, and Sinha in ICML ’20, Eurocrypt ’21, and ICML ’21.
-------------------
Other upcoming DeLTA seminars:
15 November 2021 @ 10:00. Sorawit Saengkyongam. Invariant Policy Learning: A Causal Perspective.
29 November 2021 @ 10:00. Hippolyte Bourel.
-------------------
You can subscribe to DeLTA Seminar mailing list by sending an empty email to delta-seminar-join@list.ku.dk.
Information about DeLTA Seminars is available at the DeLTA Lab page: https://sites.google.com/diku.edu/delta
DeLTA is a research group affiliated with the Department of Computer Science at the University of Copenhagen studying diverse aspects of Machine Learning Theory and its applications, including, but not limited to Reinforcement Learning, Online Learning and Bandits, PAC-Bayesian analysis